SLAC's SPINE AI Reconstructs Neutrino Collisions Particle by Particle in Giant Argon Detectors

Neutrinos pass through your body by the trillions every second without leaving a trace. Catching even a few in a detector is hard, and making sense of what they leave behind is harder. Researchers at the Department of Energy's SLAC National Accelerator Laboratory say a machine-learning program called SPINE can now do that job in a way that is both powerful and transparent.

Neutrinos are among the most abundant particles in the universe. They come from nuclear reactions in the sun, exploding stars, particle accelerators and the decay of radioactive elements on Earth. They are nearly massless and interact so weakly with ordinary matter that studying them requires enormous detectors and intense particle beams. Answers they could provide range from the nature of dark matter to why matter exists at all.

SPINE stands for Scalable Particle Imaging with Neural Embeddings. It analyzes data from liquid argon time-projection chambers, the detector type used by the Short-Baseline Neutrino program at Fermi National Accelerator Laboratory in Illinois and by the Deep Underground Neutrino Experiment, or DUNE, which is under construction in Illinois and South Dakota and expected to start taking data in 2030.

These detectors hold hundreds to thousands of tons of ultrapure liquid argon at around minus 300 degrees Fahrenheit. When a neutrino hits an argon atom, the collision can send several particles through the liquid, and each leaves a trail of electrons that drift toward collection wires. The result is a detailed three-dimensional picture. A single image can contain many overlapping neutrino interactions, with multiple tracks and electromagnetic showers mixed with cosmic rays and other background.

For years, scientists tackled this by chaining together specialized algorithms, each handling one narrow task and passing results to the next. An error at any step could ripple through the whole reconstruction. SPINE instead uses several machine-learning models that work together. They first spot basic features in the image, then identify individual particles, then piece together the complete interaction. SPINE learns from simulated detector images.

"We didn't try to go, in one step, from the raw detector data all the way to these high-level physics observables," said SLAC scientist François Drielsma, the lead developer. "We built the reconstruction hierarchically, extracting progressively higher-level information from the detector data." That also makes the program more transparent than a black-box AI model, because it shows its work by reconstructing the particles that produced the signal.

SLAC physicist Mark Convery said the effects neutrino physicists now chase are small. "They may have big physics impacts, but they're subtle," he said. "You need a really intense beam, a really capable detector and a really efficient and precise reconstruction technique. SPINE is that third leg of the stool."

The project grew out of work by SLAC scientist Kazuhiro Terao, who began applying computer-vision techniques to neutrino detector images. With DUNE years from first data, the tool is meant to be ready when the signals arrive.